def test_DIN_sum():
model_name = "DIN_sum"
x, y, feature_dim_dict, behavior_feature_list = get_xy_fd()
model = DIN(feature_dim_dict,
behavior_feature_list,
hist_len_max=4,
embedding_size=8,
use_din=False,
hidden_size=[4, 4, 4],
keep_prob=0.6,
activation="sigmoid")
model.compile('adam',
'binary_crossentropy',
metrics=['binary_crossentropy'])
model.fit(x, y, verbose=1, validation_split=0.5)
print(model_name + " test train valid pass!")
model.save_weights(model_name + '_weights.h5')
model.load_weights(model_name + '_weights.h5')
print(model_name + " test save load weight pass!")
save_model(model, model_name + '.h5')
model = load_model(model_name + '.h5', custom_objects)
print(model_name + " test save load model pass!")
print(model_name + " test pass!")
def test_DIN_att():
model_name = "DIN_att"
x, y, feature_dim_dict, behavior_feature_list = get_xy_fd()
model = DIN(
feature_dim_dict,
behavior_feature_list,
hist_len_max=4,
embedding_size=8,
use_din=True,
hidden_size=[4, 4, 4],
keep_prob=0.6,
)
model.compile('adam',
'binary_crossentropy',
metrics=['binary_crossentropy'])
model.fit(x, y, verbose=1, validation_split=0.5)
print(model_name + " test train valid pass!")
model.save_weights(model_name + '_weights.h5')
model.load_weights(model_name + '_weights.h5')
print(model_name + " test save load weight pass!")
# try:
# save_model(model, name + '.h5')
# model = load_model(name + '.h5', custom_objects)
# print(name + " test save load model pass!")
# except:
# print("【Error】There is a bug when save model use Dice---------------------------------------------------")
print(model_name + " test pass!")
def test_DIN_att():
model_name = "DIN_att"
x, y, feature_dim_dict, behavior_feature_list = get_xy_fd()
model = DIN(feature_dim_dict, behavior_feature_list, hist_len_max=4, embedding_size=8,
use_din=True, hidden_size=[4, 4, 4], keep_prob=0.6,)
model.compile('adam', 'binary_crossentropy',
metrics=['binary_crossentropy'])
model.fit(x, y, verbose=1, validation_split=0.5)
print(model_name+" test train valid pass!")
model.save_weights(model_name + '_weights.h5')
model.load_weights(model_name + '_weights.h5')
print(model_name+" test save load weight pass!")
# try:
# save_model(model, name + '.h5')
# model = load_model(name + '.h5', custom_objects)
# print(name + " test save load model pass!")
# except:
# print("【Error】There is a bug when save model use Dice---------------------------------------------------")
print(model_name + " test pass!")
def test_DIN_model_io():
model_name = "DIN_att"
_, _, feature_dim_dict, behavior_feature_list = get_xy_fd()
model = DIN(feature_dim_dict, behavior_feature_list, hist_len_max=4, embedding_size=8, att_activation=Dice,
use_din=True, hidden_size=[4, 4, 4], keep_prob=0.6,)
model.compile('adam', 'binary_crossentropy',
metrics=['binary_crossentropy'])
#model.fit(x, y, verbose=1, validation_split=0.5)
save_model(model, model_name + '.h5')
model = load_model(model_name + '.h5', custom_objects)
print(model_name + " test save load model pass!")
def test_DIN_model_io():
name = "DIN_att"
x, y, feature_dim_dict, behavior_feature_list = get_xy_fd()
model = DIN(
feature_dim_dict,
behavior_feature_list,
hist_len_max=4,
embedding_size=8,
use_din=True,
hidden_size=[4, 4, 4],
keep_prob=0.6,
)
model.compile('adam',
'binary_crossentropy',
metrics=['binary_crossentropy'])
#model.fit(x, y, verbose=1, validation_split=0.5)
save_model(model, name + '.h5')
model = load_model(name + '.h5', custom_objects)
print(name + " test save load model pass!")
def example_din():
"""
1. 生成训练数据为txt格式的，逗号分割字段
2. 转换成tfrecord
3. 读取数据，区分dense, sparse, VarLenSparse, 用户行为序列特征
4. 分别喂到模型中，看看会怎么样
:return:
"""
# x, y, feature_columns, behavior_feature_list = get_xy_fd() #说一下哪几列是当前的item需要和历史的行为进行attention的。所以之后就可以尝试，还是像之前一样读数据，然后只是把需要attention的列名单拿出来，放到list中就可以了
x, y, feature_columns, behavior_feature_list = get_xy_from_txt(
) #说一下哪几列是当前的item需要和历史的行为进行attention的。所以之后就可以尝试，还是像之前一样读数据，然后只是把需要attention的列名单拿出来，放到list中就可以了
# dataset = tf.data.Dataset.from_tensor_slices((x.values, y.values))
model = DIN(feature_columns, behavior_feature_list)
model.compile(
'adam',
keras.losses.binary_crossentropy,
metrics=[keras.metrics.AUC(), keras.metrics.categorical_accuracy])
history = model.fit(x, y, verbose=1, epochs=10, validation_split=0.5)
# history = model.fit(dataset, verbose=1, epochs=10, validation_data=(x,y))
# history = model.fit(dataset, verbose=1, epochs=10, validation_split=0.5)
print("history: ", history)
def test_DIN_sum():
model_name = "DIN_sum"
x, y, feature_dim_dict, behavior_feature_list = get_xy_fd()
model = DIN(feature_dim_dict,
behavior_feature_list,
hist_len_max=4,
embedding_size=8,
use_din=False,
hidden_size=[4, 4, 4],
keep_prob=0.6,
activation="sigmoid")
check_model(model, model_name, x, y)
def get_xy_fd():
feature_dim_dict = {"sparse": [SingleFeat('user', 3), SingleFeat(
'gender', 2), SingleFeat('item', 3+1), SingleFeat('item_gender', 2+1)], "dense": [SingleFeat('score', 0)]}
behavior_feature_list = ["item", "item_gender"]
uid = np.array([0, 1, 2])
ugender = np.array([0, 1, 0])
iid = np.array([1, 2, 3]) # 0 is mask value
igender = np.array([1, 2, 1]) # 0 is mask value
score = np.array([0.1, 0.2, 0.3])
hist_iid = np.array([[1, 2, 3, 0], [1, 2, 3, 0], [1, 2, 0, 0]])
hist_igender = np.array([[1, 1, 2, 0], [2, 1, 1, 0], [2, 1, 0, 0]])
feature_dict = {'user': uid, 'gender': ugender, 'item': iid, 'item_gender': igender,
'hist_item': hist_iid, 'hist_item_gender': hist_igender, 'score': score}
x = [feature_dict[feat.name] for feat in feature_dim_dict["sparse"]] + [feature_dict[feat.name]
for feat in feature_dim_dict["dense"]] + [feature_dict['hist_'+feat] for feat in behavior_feature_list]
y = [1, 0, 1]
return x, y, feature_dim_dict, behavior_feature_list
if __name__ == "__main__":
x, y, feature_dim_dict, behavior_feature_list = get_xy_fd()
model = DIN(feature_dim_dict, behavior_feature_list, hist_len_max=4,)
model.compile('adam', 'binary_crossentropy',
metrics=['binary_crossentropy'])
history = model.fit(x, y, verbose=1, epochs=10, validation_split=0.5)
hist_cate_id = np.array([[1, 2, 2, 0], [2, 2, 1, 0], [1, 2, 0, 0]])
seq_length = np.array([3, 3,
2]) # the actual length of the behavior sequence
feature_dict = {
'user': uid,
'gender': ugender,
'item_id': iid,
'cate_id': cate_id,
'hist_item_id': hist_iid,
'hist_cate_id': hist_cate_id,
'pay_score': pay_score,
'seq_length': seq_length
}
x = {
name: feature_dict[name]
for name in get_feature_names(feature_columns)
}
y = np.array([1, 0, 1])
return x, y, feature_columns, behavior_feature_list
if __name__ == "__main__":
x, y, feature_columns, behavior_feature_list = get_xy_fd()
model = DIN(feature_columns, behavior_feature_list)
# model = BST(feature_columns, behavior_feature_list,att_head_num=4)
model.compile('adam',
'binary_crossentropy',
metrics=['binary_crossentropy'])
history = model.fit(x, y, verbose=1, epochs=10, validation_split=0.5)
elif column == 'action_type':
feature_columns += [SparseFeat(column, 4 + 1, embedding_dim=dim)]
else:
feature_columns += [DenseFeat(column, 1)]
# maxlen为历史信息的长度，vocabulary_size为onehot的长度
feature_columns += [
VarLenSparseFeat(sparsefeat=SparseFeat('hist_merchant_id', vocabulary_size=1993, embedding_dim=8,
embedding_name='merchant_id'), maxlen=M),
VarLenSparseFeat(sparsefeat=SparseFeat('hist_action_type', vocabulary_size=4, embedding_dim=4,
embedding_name='action_type'), maxlen=M)]
history_features = ['merchant_id', 'action_type']
print(len(feature_columns))
# 使用DIN模型
model = DIN(feature_columns, history_features)
# 使用Adam优化器，二分类的交叉熵
model.compile('adam', 'binary_crossentropy', metrics=['binary_crossentropy'])
# model.compile(loss=[binary_focal_loss(alpha=.25, gamma=2)], metrics=["accuracy"])
# 组装train_model_input，得到feature names，将train_X转换为字典格式
feature_names = list(train_X.columns)
train_model_input = {name: train_X[name].values for name in get_feature_names(feature_columns)}
print("########################################")
# histroy输入必须是二维数组
from tqdm import tqdm
for fea in ['hist_merchant_id', 'hist_action_type']:
list = []
for i in tqdm(train_model_input[fea]):
sess_len_max = SESS_MAX_LEN
BATCH_SIZE = 1024
sess_feature = ['item_id']
# def auc(y_true,y_pred):
# return tf.py_func(roc_auc_score, (y_true, y_pred), tf.double)
EMBEDDING_SIZE = int(ebs)
if EMBEDDING_SIZE == 0:
EMBEDDING_SIZE = "auto"
model = DIN(fd,
sess_feature,
embedding_size=EMBEDDING_SIZE,
dnn_dropout=float(dnn_dropout),
att_activation='dice',
att_weight_normalization=False,
hist_len_max=sess_len_max,
dnn_hidden_units=(200, 80),
att_hidden_size=(
64,
16,
))
model.compile('adagrad',
'binary_crossentropy',
metrics=['binary_crossentropy'])
model_dir = "../model_dir_" + str(EMBEDDING_SIZE)
if not os.path.exists(model_dir):
os.mkdir(model_dir)
if os.path.exists(model_dir + '/ckpt.h5'):
model.load_weights(model_dir + '/ckpt.h5')
"""
test_input_pos = pd.read_pickle(
user_age = np.array([1, 2, 3])
user_gender = np.array([0, 1, 0])
item_id = np.array([0, 1, 2])
item_gender = np.array([0, 1, 0])
# multi-value feature input
hist_item_id = np.array([[0, 1, 2, 3], [0, 1, 2, 3], [0, 1, 2, 0]])
hist_item_gender = np.array([[0, 1, 0, 1], [0, 1, 1, 1], [0, 0, 1, 0]])
# valid length of behavior sequence of every sample
hist_length = np.array([4, 4, 3])
feature_dict = {'user_age': user_age, 'user_gender': user_gender, 'item_id': item_id, 'item_gender': item_gender,
'hist_item_id': hist_item_id, 'hist_item_gender': hist_item_gender, }
x = [feature_dict[feat] for feat in feature_dim_dict["sparse"]] + \
[feature_dict['hist_'+feat]
for feat in behavior_feature_list] + [hist_length]
# Notice the concatenation order: single feature + multi-value feature + length
# Since the length of the historical sequences of different features in DIN are the same(they are all extended from item_id),only one length vector is enough.
y = [1, 0, 1]
return x, y, feature_dim_dict, behavior_feature_list
if __name__ == "__main__":
x, y, feature_dim_dict, behavior_feature_list = get_xy_fd()
model = DIN(feature_dim_dict, behavior_feature_list, hist_len_max=4,)
model.compile('adam', 'binary_crossentropy',
metrics=['binary_crossentropy'])
history = model.fit(x, y, verbose=1, validation_split=0.5)
elif sys.argv[1] == 'PNN_UDG':
model = PNN_UDG(dnn_feature_columns, untrainable_features_columns, (200, 80), uid_feature_name=udg_features,
udg_embedding_size=int(sys.argv[5]))
elif sys.argv[1] == 'PNN':
model = PNN(dnn_feature_columns, untrainable_features_columns, (200, 80))
elif sys.argv[1] == 'WDL':
model = WDL(linear_feature_columns, dnn_feature_columns, [], (200, 80))
elif sys.argv[1] == 'WDL_UDG':
model = WDL_UDG(linear_feature_columns, dnn_feature_columns, untrainable_features_columns, (200, 80), uid_feature_name=udg_features, udg_embedding_size=int(sys.argv[5]))
elif sys.argv[1] == 'DIEN':
model = DIEN(fixlen_feature_columns, behavior_feature_list,
dnn_hidden_units=[200, 80], dnn_dropout=0, gru_type="AUGRU", use_negsampling=True)
elif sys.argv[1] == 'DIEN_UDG':
model = DIEN_UDG(fixlen_feature_columns, untrainable_features_columns, behavior_feature_list, dnn_hidden_units=[200, 80], dnn_dropout=0, gru_type="AUGRU", use_negsampling=True, uid_feature_name=udg_features, udg_embedding_size=int(sys.argv[5]))
elif sys.argv[1] == 'DIN':
model = DIN(fixlen_feature_columns, behavior_feature_list, dnn_hidden_units=[200, 80], dnn_dropout=0)
elif sys.argv[1] == 'DIN_UDG':
model = DIN_UDG(fixlen_feature_columns, untrainable_features_columns, behavior_feature_list, dnn_hidden_units=[200, 80], dnn_dropout=0, uid_feature_name=udg_features, udg_embedding_size=int(sys.argv[5]))
if sys.argv[4] == 'focal':
model.compile("adam", loss=focal_loss, metrics=['binary_crossentropy'], )
else:
model.compile("adam", "binary_crossentropy", metrics=['binary_crossentropy'], )
init_lr = float(tf.keras.backend.get_value(model.optimizer.learning_rate))
lr = [init_lr, init_lr/2, init_lr/4]
history_all = {}
max_auc, min_log, min_rmse, max_rig = 0, 0, 0, 0
for x in range(epoch):
tf.keras.backend.set_value(model.optimizer.lr, lr[x])
history = CustomCallback()
model.fit(train_model_input, train[target].values, batch_size=256, epochs=1, verbose=1,
test_label = label[test_idx]
sess_len_max = SESS_MAX_LEN
BATCH_SIZE = 4096
sess_feature = ['cate_id', 'brand']
TEST_BATCH_SIZE = 2**17
REG = 1e-6
model = DIN(fd,
sess_feature,
embedding_size=4,
att_activation='dice',
att_weight_normalization=False,
hist_len_max=sess_len_max,
dnn_hidden_units=(200, 80),
att_hidden_size=(
64,
16,
),
l2_reg_embedding=REG,
seed=2019)
model.compile('adagrad',
'binary_crossentropy',
metrics=[
'binary_crossentropy',
])
hist_ = model.fit(
train_input[:],
def buildModel(self):
feature_columns = self.encoder.getFeatureColumns()
self.model = DIN(feature_columns, self.encoder.behavior_list)
self.model.compile('adam',
'binary_crossentropy',
metrics=['binary_crossentropy'])
class Trainer:
def __init__(self):
self.data = None
self.encoder = None
self.model = None
# number of positive samples
self.num_pos = None
self.recipeDomain = None
def loadData(self, url: str):
self.data = pd.read_csv(url)
self.recipeDomain = pd.read_csv(url)
self.num_pos = self.data.shape[0]
self.encoder = ModelEncoder()
self.encoder.train()
def preProcessData(self):
self.data = self.encoder.encode(self.data)
self.data['result'] = [1] * self.num_pos
self.recipeDomain = self.encoder.encode(self.recipeDomain)
self.build_negative_data()
def buildModel(self):
feature_columns = self.encoder.getFeatureColumns()
self.model = DIN(feature_columns, self.encoder.behavior_list)
self.model.compile('adam',
'binary_crossentropy',
metrics=['binary_crossentropy'])
def train(self):
model_input = {}
for feat, _ in self.encoder.fixed_sparse_dict + self.encoder.var_sparse_dict:
model_input[feat] = self.data[feat]
history = self.model.fit(model_input,
self.data['gt'].values,
batch_size=256,
epochs=10,
verbose=2,
validation_split=0.2,
shuffle=True)
def dump_model(self, path: str):
save_model(self.model, path)
def dump_encoder(self, path: str):
pickle.dump(self.encoder, self, protocol=4)
def update(self):
url = ''
self.recipeDomain = pd.read_csv(url)
self.encoder.encode(self.recipeDomain)
def build_negative_data(self):
for i in range(self.num_pos):
record = self.data.iloc[i]
clicked_set = record['hist_recipe'] + record['recipe'][i]
for j in self.recipeDomain.shape[0]:
if self.recipeDomain.iloc[j]['recipe'] not in clicked_set:
# valid unclicked combination
for feat in record.columns:
if feat in self.recipeDomain.columns:
record[feat] = self.recipeDomain.iloc[j][feat]
record['result'] = 0
self.data.append(record)